Re: [PATCH v5 tip/core/locking 6/7] locking: Add an smp_mb__after_unlock_lock() for UNLOCK+LOCK barrier

["Paul E. McKenney" <paulmck@linux.vnet.ibm.com>]

On Mon, Dec 09, 2013 at 09:26:41PM -0800, Paul E. McKenney wrote:
> On Mon, Dec 09, 2013 at 05:34:17PM -0800, Josh Triplett wrote:
> > On Mon, Dec 09, 2013 at 05:28:02PM -0800, Paul E. McKenney wrote:
> > > From: "Paul E. McKenney" <paulmck@linux.vnet.ibm.com>
> > > 
> > > The Linux kernel has traditionally required that an UNLOCK+LOCK pair
> > > act as a full memory barrier when either (1) that UNLOCK+LOCK pair
> > > was executed by the same CPU or task, or (2) the same lock variable
> > > was used for the UNLOCK and LOCK.  It now seems likely that very few
> > > places in the kernel rely on this full-memory-barrier semantic, and
> > > with the advent of queued locks, providing this semantic either requires
> > > complex reasoning, or for some architectures, added overhead.
> > > 
> > > This commit therefore adds a smp_mb__after_unlock_lock(), which may be
> > > placed after a LOCK primitive to restore the full-memory-barrier semantic.
> > > All definitions are currently no-ops, but will be upgraded for some
> > > architectures when queued locks arrive.
> > > 
> > > Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
> > > Cc: Linux-Arch <linux-arch@vger.kernel.org>
> > > Cc: Ingo Molnar <mingo@redhat.com>
> > > Cc: Peter Zijlstra <peterz@infradead.org>
> > > Cc: Oleg Nesterov <oleg@redhat.com>
> > > Cc: Linus Torvalds <torvalds@linux-foundation.org>
> > 
> > It seems quite unfortunate that this isn't in some common location, and
> > then only overridden by architectures that need to do so.
> 
> I was thinking that include/asm-generic/barrier.h was the place, but
> it is all-or-nothing, used by UP architectures, from what I can see.
> I figured that if there is such a common location, posting this patch
> might flush it out.  I am not sure that this single definition is worth
> the creation of a common place -- or even this definition combined with
> smp_read_barrier_depends().

And of course the right place to put this is include/linux/spinlock.h,
the same place where smp_mb__before_spinlock() is defined.  Exceptions
then go into the corresponding arch-specific spinlock.h files.

Much better that way, thank you for calling this out!

							Thanx, Paul

> > More importantly: you document this earlier in the patch series than you
> > introduce it.
> 
> Fair point, I reversed the order of those two patches.
> 
> 							Thanx, Paul